Recognition of cough using features improved by sub-band energy transformation

The purpose of this paper is to improve mel frequency cepstrum coefficients (MFCCs) for cough recognition. To highlight high energy, the most remarkable characteristic of cough sound, we propose a method of sub-band energy transformation to improve traditional MFCCs. This method enhances bands with high energy and ignores the ones with low energy according to the sub-band energy distribution acquired by investigation of varieties of cough sounds. Cough recognition experiments using hidden Markov models (HMMs) show that the average recognition rate rises from 87% to 91% and robustness of the system in noisy environment is improved by the proposed method.